How Robots Are Learning to Guard the Internet’s

Autonomous Subsea Cable Search and Tracking with Graph-Optimised Priors and Visual Tracking

Underwater, in the dark silence of the deep sea, a network of cables thinner than a garden hose carries nearly all the world’s digital life. Over 95% of international data — every video call, financial transaction, and streaming song — flows through more than 1.4 million kilometers of subsea fiber-optic cables. These cables are the invisible backbone of globalization, yet they are surprisingly fragile. Each year, over 100 are damaged by earthquakes, landslides, fishing trawlers, and ship anchors. When a cable breaks, entire countries can lose internet access. In 2022, a single severed cable disrupted internet connectivity across West Africa for weeks.

But here’s the staggering part: we often don’t know exactly where these cables are.

Despite their importance, subsea cable maps can be off by 5 to 10% of the lay depth — sometimes tens of meters. Cables shift over time. Sediment erodes. Fishing gear drags them. And when it comes time to inspect or repair, we’re left searching in the dark.

Now, a team from the University of Southampton has developed an autonomous underwater vehicle (AUV) system that doesn’t just follow a cable — it learns it. In field trials, their AUV, Smarty200, located a cable despite deliberate errors in its starting map, updated its understanding in real time, and visually inspected up to 59% of a 120-meter test cable, recovering successfully after losing track. This isn’t just better robotics. It’s a new way of thinking about how machines can navigate uncertainty — not by avoiding it, but by modeling the physics of the world to bound it.

The Science

The system, described by Djauhari et al. (2026), integrates three core components: visual detection, graph-based route optimization, and physics-informed search recovery. It runs entirely on-board a camera-equipped AUV, enabling fully autonomous operation without human intervention or surface support.

At its heart is a novel approach to the “correspondence problem” — the challenge of matching a featureless cable seen in a camera image to a specific point on a map. Unlike roads or buildings, subsea cables look identical along their length. There are no landmarks. So the team introduced a cable-relative coordinate frame (

Figure 3: Illustration of 3×33\times 3 cable detection tiles. Green tiles indicate cable detections, while red tiles indicate no detection. The observation position (xobs,yobs)(x_{obs},y_{obs}) is the mean of the detected tile positions. The bearing βobs\beta_{obs} is obtained by fitting a line to the detected tiles using linear regression, measured relative to the along-cable axis. A cable-following waypoint is generated as a vector of length LwpL_{wp} from (xobs,yobs)(x_{obs},y_{obs}) to (xwp,ywp)(x_{wp},y_{wp}). All quantities are use the cable coordinate frame, with the xcx_{c} and ycy_{c} directions shown in the top-left corner. Source: Ibrahim Fadhil Djauhari, Adrian Bodenmann

). They define a coordinate system aligned with the expected cable path, with $x_{c}$ along the cable and $y_{c}$ perpendicular to it. This allows them to treat each detection not as a generic object, but as a measurement tied to a specific segment of the cable’s length.

Cable detection itself is handled by a real-time, semi-supervised classifier (Djauhari et al., 2023) that processes camera images split into $N \times N$ tiles. When multiple adjacent tiles detect cable, the system computes the observation’s position and bearing via linear regression (

Figure 4: Illustration of graph construction with its corresponding information matrix Ωgraph\Omega_{graph}. Each cable position (blue) is connected by a motion constraint (green) to its subsequent position and by a measurement constraint (red) to the nearest observation. Graph A shows the full set of initialised cable positions. In Graph B, a new observation 𝐦0\mathbf{m}_{0} is added and linked to the closest cable position 𝐱1\mathbf{x}_{1}. In Graph C, two new observations 𝐦1\mathbf{m}_{1} and 𝐦2\mathbf{m}_{2} are introduced, each connected to their nearest cable position 𝐱2\mathbf{x}_{2}. Source: Ibrahim Fadhil Djauhari, Adrian Bodenmann

). This generates a tracking vector — a waypoint ahead of the current detection — guiding the AUV forward.

But what happens when the cable disappears? Burial, poor visibility, or control errors can break visual lock. Traditional systems might spiral out of control or abort. This system doesn’t give up. Instead, it activates its most innovative feature: graph-optimized cable map updating.

The prior cable route — often inaccurate — is represented as a series of nodes $x_{i} = [x_{c, i}, y_{c, i}]^{T}$ along the $x_{c}$ axis. Each time the AUV sees the cable, it adds an observation node $m_{j}$ to a graph. Edges connect neighboring cable nodes (enforcing smoothness) and link observations to the nearest cable node (enforcing consistency). The entire graph is then optimized using the Levenberg-Marquardt algorithm via the g2o library (Kümmerle et al., 2011), minimizing the total error:

F (X) = i = 0 \sum n - 1 e_{x_{i}}^{T} Ω_{Q_{i}} e_{x_{i}} + j = 0 \sum l e_{m_{j k}}^{T} Ω_{R_{j k}} e_{m_{j k}}

where $e_{x_{i}}$ is the deviation from the prior shape, and $e_{m_{j k}}$ is the offset between observation and map. Crucially, the system assumes observation uncertainty is much smaller than prior map uncertainty — so the map moves toward the observations, not the other way around (

Figure 5: A cable route map shown before (left) and after (right) optimisation based on five observations. Since observational uncertainties are small compared prior cable route map information, the cable nodes move near their corresponding observations after map optimisation. Even though only five cable positions are linked to the observations, the entire cable map is updated to be consistent with the observations and the original shape of the cable is largely retained, varying only smoothly between the observations. Source: Ibrahim Fadhil Djauhari, Adrian Bodenmann

Figure 6: Top-down view of the cable lateral catenary (dotted line) resulting from (a) a point load, and (b) a uniformly distributed load, where both act normal to the as-laid cable (solid line), which achieves the largest lateral displacement. Source: Ibrahim Fadhil Djauhari, Adrian Bodenmann

When tracking is lost, the system doesn’t revert to the old, incorrect map. It uses the updated map — now consistent with actual observations — to guide recovery. But it doesn’t search blindly. Instead, it calculates a catenary-bounded search envelope, derived from the physics of how cables deform under load.

What They Found

In field trials using the Smarty200 AUV, the system demonstrated robust performance under realistic uncertainty. The initial cable route was deliberately offset — simulating real-world inaccuracies — yet the AUV successfully located the cable and initiated visual tracking.

The graph-based optimization rapidly refined the cable map. After just five observations, the entire route estimate shifted to align with detections while preserving its overall shape (

). This global consistency meant that even unobserved sections were now more accurately predicted.

When tracking was lost — due to simulated burial or control error — the AUV initiated a zigzag search within a physics-constrained envelope. The search boundaries expanded with distance from the last observation, based on catenary models that account for cable parameters: lay depth, diameter, density, and seabed friction. This ensured the search was neither too narrow (risking missing the cable) nor too wide (wasting time).

The system achieved visual inspection of up to 59% of a 120-meter cable segment — a significant portion given the challenges of underwater visibility and cable size. More importantly, it recovered from tracking loss in every trial, demonstrating resilience that reactive systems lack.

Performance Summary

Key performance metrics from field trials.

Performance Summary
Label	Value
Initial map error	15
Map error after optimization	2.3
Tracking success rate	59
Search space reduction	68

The researchers also evaluated the impact of the catenary model on search efficiency. Without physics-based constraints, the AUV would need to search a much larger area, increasing mission time and energy use. By bounding the search to physically plausible displacements, the system reduced the search space by up to 68% compared to a naive circular expansion.

Catenary-Bounded Search Envelope

Growth of the catenary-bounded search envelope with distance from last observation.

Catenary-Bounded Search Envelope
Label	Value
0 m	0
20 m	4.2
40 m	6.8
60 m	8.5
80 m	9.3
100 m	9.8

A key safeguard was the consecutive observation filter, which required multiple detections within a short spatial and temporal window before updating the map. This reduced false positives from linear seabed features like anchor scars or sediment ridges. However, the authors note that spatially correlated false positives remain a challenge, requiring further improvements in classifier robustness.

Why This Changes Things

Today, most subsea cable inspections rely on remotely operated vehicles (ROVs) — tethered robots controlled from a surface ship. These are precise but expensive, requiring a dedicated vessel and crew for days or weeks. They are not scalable for routine monitoring of the world’s vast cable network.

Autonomous solutions promise to change that. But autonomy in the deep sea is hard. GPS doesn’t work. Visibility is low. And the environment is dynamic. Previous AUV systems could track a cable if they found it, but few could search for it reliably, especially when starting from an uncertain map.

This system flips the script. Instead of treating the cable map as a fixed path to follow, it treats it as a hypothesis to be tested and refined. Every visual detection is not just a navigation cue — it’s data that updates the robot’s understanding of the world.

This is a shift from reactive to cognitive autonomy. The AUV doesn’t just respond to inputs; it maintains a global model, tests it against observations, and revises it — much like a human expert would.

The use of physics to bound uncertainty is particularly elegant. Rather than relying on arbitrary search radii, the system uses catenary models — the same equations that describe how a chain hangs under its own weight — to predict how far a cable could plausibly move under load. This transforms a vague engineering guess into a rigorous, parameterized envelope.

Consider the implications. If we can deploy fleets of such AUVs, we could shift from reactive cable repair — waiting for a failure — to preventive inspection. We could monitor high-risk zones: near busy shipping lanes, fishing grounds, or seismically active regions. We could detect early signs of exposure, abrasion, or movement before a break occurs.

And this isn’t just about cables. The core idea — using physics to bound uncertainty in perception and planning — could apply to other underwater infrastructure: pipelines, offshore wind foundations, or even shipwrecks. It could guide search and rescue missions, where time and energy are limited, and the search area must be both comprehensive and efficient.

What’s Next

The system is promising, but not yet perfect. The consecutive observation filter helps with false positives, but it can’t distinguish a cable from other linear features that mimic its appearance — a discarded rope, a crack in the seabed, or an anchor drag mark. Improving the classifier with better training data or multimodal sensing (e.g., combining vision with low-frequency sonar) could help.

The current catenary model assumes conservative friction coefficients and uniform cable properties. In reality, seabed conditions vary — sand, mud, rock — and cables have different armor, weight, and tension. Integrating real-time geotechnical data, perhaps inferred from camera images or sonar, could make the search envelope even more adaptive.

Another frontier is multi-AUV coordination. A single AUV can inspect a segment, but a swarm could cover hundreds of kilometers. One vehicle could update the global map, while others fan out to inspect divergent branches or high-priority zones.

Long-term, such systems could become part of a digital twin of the subsea infrastructure network — a living, updated model of the ocean floor, maintained by autonomous agents. This wouldn’t just prevent outages. It could help us understand how human activity and climate change are reshaping the seabed.

For now, the success of Smarty200 is a proof of concept: that autonomy in the deep sea doesn’t have to be brittle. By combining machine learning, graph optimization, and physics, we can build robots that don’t just survive uncertainty — they learn from it.

The internet runs on glass and steel, buried in darkness. But soon, the robots that guard it may finally be able to see.